Announcing our new Foundation for Deep Learning acceleration  MIOpen 1.0 which introduces support for Convolution Neural Network acceleration — built to run on top of the ROCm software stack!

This release includes the following:

  • Deep Convolution Solvers  optimized for both forward and backward propagation
    • Optimized Convolutions including Winograd and FFT transformations
    • Optimized GEMM’s for Deep Learning
    • Pooling, Softmax, Activations, Gradient Algorithms Batch Normalization, and LR Normalization
    • MIOpen describes data as 4-D tensors ‒ Tensors 4D NCHW format
  • Support for OpenCL and HIP enabled frameworks
  • MIOpen Driver enables to testing forward/backward network of any particular layer in MIOpen.
  • Binary Package support for Ubuntu  16.04 and Fedora 24
  • Source code at ROCm Software Platform Github site

Quickstart Instructions to get familiar with MIOpen via the MIOpen Driver

Here’s a simple workflow to get you quickly up and running with OpenCL on ROCm —

Install the MIOpen implementation (assuming you already have the ‘rocm’  and ‘rocm-opencl-dev” package installed) and Repo Server setup on your box:

For just OpenCL development  
    sudo apt-get install miopengemm miopen-opencl 
For HIP development
    sudo apt-get install miopengemm miopen-hip